Deep learning has great potential in the field of functional peptide prediction. This study combines metagenomics and deep learning to efficiently discover potential umami peptides in fermented sausages. A candidate peptide library was generated using metagenomic data from fermented sausages, an integrated deep learning model was constructed for prediction, and SHAP (SHapley Additive exPlanations) interpretability analysis was performed to elucidate the key amino acid features and contributions of the model in predicting umami peptides, screening the top ten peptides with the highest predicted probability. Subsequently, molecular docking was performed to assess the binding stability of these peptides with the umami receptor T1R1/T1R3, selecting the three peptides DDSMAATGL, DGEEDASM, and DEEEVDI with the most stable binding for further study. Docking analysis revealed the important roles of the key receptor residues Glu301, Arg277, Lys328, and His71 in hydrogen bond formation. Molecular dynamics simulations validated the robust integrity of the peptide–receptor associations. Finally, sensory evaluation demonstrated that these three peptides possessed significant umami characteristics, with low umami thresholds (0.11, 0.37, and 0.44 mg/mL, respectively). This study, based on metagenomics and deep learning, provides a high-throughput strategy for the discovery and validation of functional peptides.